Multi-modal data fusion driven spatiotemporal dynamic carbon emission accounting method and system

Abstract

The application discloses a multi-modal data fusion driven space-time dynamic carbon emission accounting method and system, and the method comprises the following steps: obtaining original emission inventory data and a plurality of proxy variables of a research area; constructing a multi-modal data fusion driven carbon emission total amount model; the multi-modal data fusion driven carbon emission total amount model comprises a rule type distributor and a learning type distributor, the rule type distributor is used for constructing initial weights inside a list grid unit according to the proxy variables; the learning type distributor is used for learning a share structure inside the list grid unit from the proxy variables of a multi-channel space according to the initial weights; according to the original emission inventory data and the proxy variables, the list grid total amount is distributed based on the multi-modal data fusion driven carbon emission total amount model, and high-resolution results are obtained. The application can effectively identify problems such as grid internal emission difference, emission hotspot position ambiguity and unclear emission gradient.

Description

Technical Field

[0001] This invention relates to the field of carbon emission accounting technology, specifically to a spatiotemporal dynamic carbon emission accounting method and system driven by multimodal data fusion. Background Technology

[0002] Under the dual-carbon goal, carbon emission accounting is gradually shifting from national and provincial total statistics to refined identification of urban agglomerations, intra-city areas, and key emission units. For multi-center urban agglomerations like the Chengdu-Chongqing Economic Circle, industrial parks, transportation corridors, urban built-up areas, and mountainous ecological spaces are interspersed, resulting in significant spatial heterogeneity in emission activities. Traditional administrative unit statistics can support macro-level total assessments, but they struggle to reveal the specific locations of emission hotspots, emission gradients, and sectoral differences within the urban agglomeration. As the requirements for detailed emission information continue to increase in coordinated pollution reduction and carbon reduction governance, refined carbon verification, and urban spatial management, constructing a kilometer-level carbon emission spatial base map has become a crucial foundation for regional carbon governance to move from total quantity control to spatial implementation.

[0003] Global and regional carbon emission inventories provide a crucial data foundation for cross-regional comparisons, long-term trend analysis, and large-scale carbon budget assessments, enabling anthropogenic emissions to be incorporated into air quality simulations, carbon budget analysis, and policy evaluations in a grid-like format. However, applications within urban clusters typically require more granular spatial representations. Pollution exposure assessments, identification of key emission areas, intra-city emission reduction zoning, and emission analysis along transportation corridors all rely on spatial inputs at the kilometer level or even finer scales, while most global or regional inventories still struggle to directly represent local differences within inventory grid units. Current representative greenhouse gas and anthropogenic emission inventories cover multiple scales, including global, regional, and urban levels.

[0004] In the process of gridding and downscaling emissions inventories, the basic practices of spatial allocation are similar. Most products use proxy variables to allocate total emissions from a region or sector to a finer grid. This type of allocation is usually based on the statistical correspondence between proxy variables and emission sources, rather than directly expressing the emission formation mechanisms of each sector. General proxies such as population density can characterize the intensity of human activity, but they are difficult to distinguish between emission differences between industrial and residential areas within a city. Using ODIAC (Open-source Data Inventory for Anthropogenic) as an example... Global fossil fuels Taking a high-resolution emissions dataset as an example, this list will include global fossil fuels. Emissions data have been extended to 1 km and incorporate point source databases and nighttime light information. However, non-point source emissions still heavily rely on nighttime light signals, and the nighttime light saturation effect can easily blur the gradient between high- and low-emission areas. Regional inventories such as CAMS-REG-ANT (CAMS Regional Anthropogenic Emissions) and MEIC (Multi-resolution Emission Inventory model for Climate and air pollution research) offer more detailed sector and fuel matching, but existing research has shown that spatial allocation errors can be amplified with increasing grid resolution. Therefore, the main problem with existing products is not a lack of inventory foundation, but rather that the spatial allocation within inventory grid cells is highly dependent on proxy variables. Simply increasing grid resolution does not automatically improve allocation quality; the key lies in whether the proxy variables match the emission mechanisms of each sector.

[0005] Spatial proxy weighting is the most commonly used method for gridding and downscaling of emission inventories. Early studies often used single GIS area source allocation, distributing regional totals to spatial units based on statistical indicators such as land use and population. Subsequently, land cover, population density, transportation networks, nighttime light, and point source information have gradually been incorporated into multi-proxy combination systems to enhance the characterization of spatial heterogeneity among different emission sources. However, the effectiveness of such methods still depends on whether the proxy variables can correspond to the spatial formation mechanisms of different sectors. When the proxy variables mainly reflect the intensity of human activities and cannot distinguish the differences between sectors such as industry, transportation, and residential, spatial bias at the fine grid unit scale is still difficult to avoid.

[0006] In the spatial redistribution process, the conservation of total quantities also affects the interpretability of downscaling results. Tobler, in proposing Pycnophylactic interpolation, emphasized that regional total quantities should remain unchanged before and after interpolation. For emission inventories, coarse-resolution inventories typically serve as total quantity constraints, while fine-grid cell results must restore internal spatial heterogeneity while maintaining consistency in total quantities across the inventory grids. Existing schemes often treat conservation constraints as ex-post verification of the results, failing to fully incorporate them into the internal mechanisms of the allocation process. In high-resolution emission estimation, spatial allocation depends not only on the reasonableness of proxy variables but also on whether the fine-grid results can maintain the total quantity relationship with the original inventory. Otherwise, it will be difficult to distinguish whether differences between different products stem from total quantity inputs, proxy variable selection, or the spatial allocation method itself.

[0007] To alleviate the difficulty of characterizing sectoral emission differences using single proxy variables, recent research has gradually shifted towards multi-source spatial data fusion. Population, land use, road network, nighttime light, point source information, and vegetation indices are currently common spatial proxy variables, providing complementary information on social activities, economic production, and ecological context: population and points of interest data reflect resident activities and urban functional agglomeration; road, nighttime light, land use, and point source information characterize transportation corridors, construction intensity, and industrial space; and ecological context information such as vegetation indices helps identify areas with low construction intensity or low human activity. However, the emission formation mechanisms of different sectors differ significantly. Transportation emissions are typically distributed along road corridors, and road networks reflect spatial patterns better than population density; residential combustion sources are constrained by population, building density, and residential space distribution; and industrial emissions are more driven by point source location, industrial land use, and nighttime light signals. Therefore, the key to multimodal carbon emission data fusion is not merely increasing the types of data, but establishing a matching relationship between sectoral emission mechanisms and proxy variables.

[0008] Regarding the verification of emission downscaling results, fine-scale... Emissions estimation has long faced the problem of insufficient reference fields. Chemical transport models can indirectly verify the rationality of spatial allocation by comparing simulated concentrations with ground monitoring data, but simulation errors are simultaneously affected by meteorological fields, chemical mechanisms, and boundary conditions, making it difficult to attribute deviations solely to emission inputs. City-based bottom-up inventories can serve as another type of reference, but comparisons between inventories can only reveal relative differences between products, making it difficult to determine which set is closer to actual emissions. Satellite inversion estimates emissions from atmospheric concentrations, while providing an independent verification path, is still constrained by transport models, prior inventories, and column concentration sensitivity. Existing uncertainty studies further indicate that the spatialization process itself is a significant source of error in high-resolution products. In the absence of publicly available 1 km data... Under the condition of a reference field, conducting cross-comparisons using multiple emission inventories with independent methodologies and applying them by introducing another type of emission inventory in the same region is a feasible approach to assess the stability and applicability of downscaling results.

[0009] In view of the above, this application is hereby submitted. Summary of the Invention

[0010] The purpose of this invention is to provide a spatiotemporal dynamic carbon emission accounting method and system driven by multimodal data fusion, which solves the problems of existing coarse-resolution methods. When global or regional carbon emission inventories are applied at regional scales such as urban agglomerations, the mixing of multiple emission sources within a single grid cell—including urban built-up areas, transportation arteries, industrial land, and agricultural and forestry spaces—leads to technical challenges such as the inability to effectively identify emission differences within the grid, unclear emission hotspot locations, and ambiguous emission gradients. This paper also addresses how to accurately reproduce the spatial heterogeneity of emissions within the grid caused by different human activities and land use types during the downscaling process from low-resolution emission data to high-resolution fine grids, while strictly ensuring that the total emissions of each coarse grid cell remain constant before and after downscaling, avoiding the introduction of new allocation errors. Furthermore, it addresses how to effectively integrate multimodal, multi-resolution, and multi-temporal spatial proxy variables such as nighttime light, road networks, land use, population, POI, and vegetation indices within a unified grid framework, enabling them to collaboratively characterize the spatial formation mechanisms of different emission sectors (area sources, line sources, and point sources) within complex urban agglomerations.

[0011] This invention is achieved through the following technical solution:

[0012] In a first aspect, the present invention provides a spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion, the method comprising:

[0013] The study area's raw emission inventory data and multiple proxy variables were obtained; the raw emission inventory data consisted of multi-source, low-resolution data.

[0014] A multimodal data fusion-driven total carbon emission model is constructed. The multimodal data fusion-driven total carbon emission model includes a rule-based allocator and a learning allocator. The rule-based allocator is used to construct the initial weights within the list grid cells based on surrogate variables. The learning allocator is used to learn the share structure within the list grid cells from surrogate variables in the multi-channel space based on the initial weights.

[0015] Based on the original emission inventory data and proxy variables, a total carbon emission allocation model driven by multimodal data fusion is used to allocate total emissions across the inventory grid, resulting in high-resolution results.

[0016] Furthermore, proxy variables include roads, nighttime lighting, built-up land, population, POI, and NDVI.

[0017] Furthermore, the rule-based allocator is an adaptive fusion coefficient based on the non-zero coverage of the proxy variable, which enables the weight allocation to be dynamically adjusted according to the proxy variable in different coarse grid cells to obtain the initial weight;

[0018] The learning-based allocator employs a teacher-guided, weakly supervised learning strategy, using initial weights as teacher signals for training.

[0019] Furthermore, based on the adaptive fusion coefficients with non-zero coverage of the proxy variables, the weight allocation can be dynamically adjusted according to the proxy variables within different coarse grid cells to obtain initial weights, including:

[0020] Based on the proxy variables, calculate the non-zero coverage of the first and second proxy variables within the coarse grid cells;

[0021] Based on the non-zero coverage of the first and second proxy variables, construct adaptive fusion coefficients;

[0022] The fused signal of the two is calculated based on the adaptive fusion coefficient and the corresponding fine mesh cell proxy signal;

[0023] The initial weights are calculated based on the fused signal and the third proxy variable.

[0024] Furthermore, the expression for the initial weights is:

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] In the formula, As the initial weights, For signal fusion, For the third surrogate variable share of fine grid cell i, The enhancement coefficient for the third proxy variable;

[0030] For fine mesh element i, the coarse mesh element belongs to The fusion coefficient, i.e., the adaptive fusion coefficient; For the first surrogate variable signal of fine mesh element i, This is the second proxy variable signal for the fine mesh element i;

[0031] For adaptive fusion coefficients, This indicates the non-zero coverage of the proxy variable within the coarse grid cell c. `clip(·)` restricts the adaptive fusion coefficients to a given interval. for proxy variables Non-zero coverage within coarse grid cell c, for proxy variables Non-zero coverage within coarse grid cell c; This is the lower bound of the adaptive fusion coefficient. This represents the upper limit of the adaptive fusion coefficient;

[0032] For non-zero coverage, This is the set of effective fine grid cells within coarse grid cell c after effective masking; the superscript v indicates the effective state. For indicator functions, Let z be the value of the surrogate variable on the fine mesh cell i.

[0033] Furthermore, after obtaining the initial weights, the process also includes:

[0034] The low vegetation cover index derived from the fourth and fifth proxy variables constitutes a backup signal; the initial weights are optimized and allocated based on the backup signal to obtain effective weights.

[0035] The expression for effective weights is:

[0036] ;

[0037] in, For effective weights, As the initial weights, This is the scaling factor for the backup signal. It is a very small positive number. As a backup signal, For fine mesh element i, the coarse mesh element belongs to The sum of the initial weights. For fine mesh element i, the coarse mesh element belongs to The summation of internal backup signals.

[0038] Furthermore, the teacher share in teacher-guided, weakly supervised learning strategies is:

[0039] ;

[0040] In the formula, For list grid cells The proportion of regular teachers in inner fine grid cell i. For the initial weights Effective weights are formed using the aforementioned effective weight expression. Then, the regular emission amount of fine grid cell i is obtained by normalizing the total amount of coarse grid according to the conservation law. The emission rate is the regular emission rate of fine grid cell j within the same coarse grid cell.

[0041] Furthermore, the original emission inventory data includes the Global Atmospheric Research Emissions Database (EDGAR), the Global Gridded Daily Carbon Dioxide Emissions Dataset (GRACED), and the Integrated Emission Inventory of Air Pollutants in China (INTAC).

[0042] Secondly, this invention also provides a spatiotemporal dynamic carbon emission accounting system driven by multimodal data fusion, the system comprising:

[0043] The acquisition unit is used to acquire raw emission inventory data and multiple proxy variables for the study area; the raw emission inventory data is multi-source low-resolution data.

[0044] The model building unit is used to build a carbon emission total model driven by multimodal data fusion. The carbon emission total model driven by multimodal data fusion includes a rule-based allocator and a learning allocator. The rule-based allocator is used to construct the initial weights inside the list grid cell based on the proxy variables. The learning allocator is used to learn the share structure inside the list grid cell from the proxy variables in the multi-channel space based on the initial weights.

[0045] The grid allocation unit is used to allocate the total carbon emission amount in the inventory grid based on the original emission inventory data and proxy variables, using a multimodal data fusion-driven total carbon emission model, to obtain high-resolution results.

[0046] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described multimodal data fusion-driven spatiotemporal dynamic carbon emission accounting method.

[0047] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0048] 1. The core advantage of this invention lies in its conservation computational structure of weighting followed by normalization. Whether using a rule-based or learning-based allocator, the generated weights are normalized within each coarse grid cell before the total coarse grid amount is precisely allocated to each fine grid cell. This mechanism, from an algorithmic structure perspective, guarantees absolute conservation of the total amount at the coarse grid scale between the downscaling result and the original inventory, avoiding uncertainties caused by post-correction and ensuring the physical consistency of the data product. This invention can effectively identify technical problems such as emission differences within the grid, ambiguous emission hotspot locations, and unclear emission gradients.

[0049] 2. This invention innovatively integrates six types of proxy variables—roads, nighttime light, built-up land, population, POI, and NDVI—surpassing traditional single or bivariate methods. It comprehensively characterizes the spatial tendency of emissions from multiple dimensions, including activity intensity, transportation corridors, and land surface functions. An adaptive fusion coefficient based on the non-zero coverage of proxy variables is designed, allowing weight allocation to be dynamically adjusted according to the effectiveness of proxy variables in different coarse grid cells, avoiding spatial structure distortion caused by fixed weights. Addressing the industry pain point of lacking fine-scale carbon emission reference data, a weakly supervised "teacher-student" learning strategy is adopted, using rule-based results as teacher signals for training, effectively solving the problem of lacking ground truth values ​​in deep learning model training.

[0050] 3. The framework design of this invention is universal, and its successful application to different sources has been demonstrated in experiments. List (EDGAR vs. GRACED) and another category of pollutants The method (INTAC) can stably reconstruct the main emission spatial pattern of the urban agglomeration's dual core-corridor-periphery under different inputs, proving that the method does not depend on specific input data and has good potential for cross-inventory and cross-pollutant migration applications. Attached Figure Description

[0051] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 The flowchart shows the spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion according to the present invention.

[0053] Figure 2 This is a framework diagram of the spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion of the present invention;

[0054] Figure 3 This is a schematic diagram of the research area and data of this invention;

[0055] Figure 4 For the EDGAR input of this invention Downscaling results;

[0056] Figure 5 For the GRACED input of this invention Downscaling results;

[0057] Figure 6 This is a structural block diagram of the spatiotemporal dynamic carbon emission accounting system driven by multimodal data fusion according to the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0059] All abbreviations are explained here:

[0060] EDGAR (Emissions Database for Global Atmospheric Research): A global database of emissions for atmospheric research. Developed under the leadership of the Joint Research Centre (JRC) of the European Commission, it is one of the most widely used gridded databases of anthropogenic greenhouse gas and atmospheric pollutant emissions globally.

[0061] GRACED (Global gRidded dAily Emissions Dataset: A global gridded daily carbon dioxide emissions dataset. Developed by Tsinghua University's Liu Zhu team in collaboration with international institutions, it is a representative near real-time, daily dataset. A resolution-based dataset of human-generated CO2 emissions.

[0062] INTAC (Integrated Emission Inventory of Air Pollutants for China): A high-resolution integrated emission inventory of air pollutants for China. Developed by the Tsinghua University MEIC team in collaboration with Peking University, Nanjing University, Jinan University, and other institutions, it is a domestically developed Chinese system. ( ), 2017, monthly, multi-source integrated authoritative grid-based emission inventory of air pollutants.

[0063] ODIAC (Open-source Data Inventory for Anthropogenic Global fossil fuels High-resolution emissions dataset. Developed by the National Institute for the Environment (NIES) of Japan under the GOSAT greenhouse gas satellite project, it is a representative 1 km resolution, monthly, global fossil fuel dataset. Emissions dataset.

[0064] CAMS-REG-ANT (CAMS Regional Anthropogenic Emissions): The CAMS Regional Anthropogenic Emissions Inventory. It is a high-resolution anthropogenic emissions inventory for the European region under the EU's Copernicus Atmosphere Monitoring Service (CAMS), developed primarily by the Dutch TNO, and designed for European air quality simulation and policy assessment.

[0065] MEIC (Multi-resolution Emission Inventory model for Climate and airpollution research): A high-resolution anthropogenic emission inventory model independently developed by Tsinghua University since 2010. Its core advantage is its high accuracy and localized emission factors for the Chinese region, making it a widely used dataset in domestic emission research.

[0066] This invention presents a spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion, which designs a carbon emission total conservation downscaling method. This is based on the consideration that... While emission inventories can constrain the total amount of emissions across the grid, they cannot directly provide the emission share of different 1 km finer grid cells within the grid. This invention aims to restore the fine-scale spatial structure while maintaining the total amount of emissions across the grid. It constructs a multimodal data fusion-driven carbon emission totals model (MCSD).

[0067] The multimodal data fusion-driven total carbon emission model (MCSD) first uses multi-source spatial proxy variables to characterize the relative emission propensity of 1 km fine-grid cells, and then in each... Within each list grid cell, these tendencies are normalized into shares, which are then used to allocate the original list grid totals. Proxy variables only affect the relative allocation within the same list grid cell and do not change the given list grid totals; therefore, the downscaling results satisfy total conservation within each list grid cell. The entire model framework comprises two paths: a regular allocator and a learning allocator. The regular allocator constructs initial weights based on proxy variables for roads, nighttime lights, construction land, and reserves; the learning allocator learns the share structure within list grid cells from multi-channel spatial proxy variables. Both paths share the same normalization backend within list grid cells, and the conservation relationship is directly guaranteed by the computational structure.

[0068] Let C be the set of coarse grid cells, and I be the set of fine grid cells (1 km apart). For any fine grid cell... Let g(i) ∈ C be the coarse grid cell to which it belongs; let g(i) be the set of effective 1km grid cells within coarse grid cell c. The total list size of coarse grid cell c is Ec, and the downscaling emission of fine grid cell i is e. i Regardless of whether a rule-based or learning-based path is used, the front-end first constructs non-negative weights for each fine-grid cell, and the back-end then normalizes these weights into shares within the corresponding list of grid cells. The effective weights after processing with the backup rules are denoted as... i, then:

[0069] (1)

[0070] From Equation 1, we can obtain:

[0071] (2)

[0072] Therefore, the downscaling results maintain the original inventory grid total within each coarse grid cell. (Subsequent...) Different allocators only change The way i is generated does not change the conservation structure defined by Equations 1 and 2.

[0073] The above markings apply to a single emission category. In In the test, the coarse-resolution emission inventory provides emissions by sector, therefore a subscript q is introduced to mark the emission category: the total amount of coarse grid cell c in category q is denoted as... The corresponding share, weight, and fine grid cell emissions are denoted as follows: , i,q and . In the experiment, q only corresponds to the total of all departments, and this subscript does not mean that the list has departmental breakdown information.

[0074] Example 1

[0075] like Figure 1 and Figure 2 As shown, the present invention provides a spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion, which includes:

[0076] Step 1: Obtain the raw emission inventory data and multiple proxy variables for the study area; the raw emission inventory data is multi-source low-resolution data.

[0077] Step 2: Construct a carbon emission total model driven by multimodal data fusion. The carbon emission total model driven by multimodal data fusion includes a rule-based allocator and a learning allocator. The rule-based allocator is used to construct the initial weights within the list grid cells based on the proxy variables. The learning allocator is used to learn the share structure within the list grid cells from the proxy variables in the multi-channel space based on the initial weights.

[0078] Step 3: Based on the original emission inventory data and proxy variables, perform total emission allocation for the inventory grid using a carbon emission total model driven by multimodal data fusion to obtain high-resolution results.

[0079] In this embodiment, the original emission inventory data includes the Global Atmospheric Research Emissions Database (EDGAR), the Global Gridded Daily Carbon Dioxide Emissions Dataset (GRACED), and the Integrated Emission Inventory of Air Pollutants in China (INTAC).

[0080] In this embodiment, proxy variables include roads, nighttime light, construction land, population, POI, and NDVI. POI refers to Point of Interest, and NDVI refers to Normalized Difference Vegetation Index.

[0081] In this embodiment, the rule-based allocator is an adaptive fusion coefficient based on the non-zero coverage of the proxy variable, which enables the weight allocation to be dynamically adjusted according to the proxy variable in different coarse grid cells to obtain the initial weight.

[0082] The learning-based allocator employs a teacher-guided, weakly supervised learning strategy, using initial weights as teacher signals for training.

[0083] Specifically, the regular allocator is denoted as MCSD-R. Its function is to convert multi-source spatial surrogate variables into relative weights within the list grid cells. Let... This represents the road signal of fine grid cell i. This represents nighttime light signals. Roads reflect traffic corridors and linear construction activities, while nighttime lights reflect the intensity of activity in built-up areas; the effective coverage of the two is not consistent in different coarse grid cells, therefore it is not advisable to use a fixed-ratio fusion throughout the study area.

[0084] First, calculate the non-zero coverage of the first surrogate variable (road) and the second surrogate variable (night light) within the coarse grid cells; for any non-negative surrogate variable... Its non-zero coverage within the coarse grid cell c is defined as:

[0085] (3)

[0086] in, For non-zero coverage, This is the set of effective fine grid cells within coarse grid cell c after effective masking; the superscript v indicates the effective state. This is an indicator function. The larger the value, the more effective fine grid cells the surrogate variable covers within the coarse grid cells.

[0087] Secondly, based on the non-zero coverage of the first proxy variable (road) and the second proxy variable (night light), adaptive fusion coefficients are constructed;

[0088] (4)

[0089] in, For adaptive fusion coefficients, This represents the non-zero coverage of the surrogate variable within the coarse grid cell c. for proxy variables Non-zero coverage within coarse grid cell c, for proxy variables Non-zero coverage within coarse grid cell c; This is the lower bound of the adaptive fusion coefficient. This is the upper limit of the adaptive fusion coefficients; clip(·) restricts the adaptive fusion coefficients to a given interval, taking the value when the denominator is 0. = 0.5; This processing avoids a single proxy variable from completely dominating the weight allocation due to differences in local coverage.

[0090] Furthermore, based on the adaptive fusion coefficients and the corresponding fine-mesh cell proxy signals, the fused signal of the two is calculated, namely the fused signal of the road and the night light, denoted as:

[0091] (5)

[0092] in, For signal fusion, is the fusion coefficient of the coarse mesh element g(i) to which the fine mesh element i belongs, i.e., the adaptive fusion coefficient; For the road signal of fine grid cell i, This represents the luminous signal of fine-grid cell i.

[0093] Then, in In the experiment, the share of construction land area was further used to enhance the basic fusion signal. Let β i Let i be the share of construction land in fine grid cell i, then the initial weight is defined as:

[0094] (6)

[0095] In the formula, As the initial weights, For signal fusion, The third proxy variable share (i.e., construction land share) for fine grid cell i. This is the enhancement coefficient for the third proxy variable (i.e., the enhancement coefficient for construction land).

[0096] As a further implementation, some coarse grid cells may have all-zero rule weights; direct normalization would prevent these coarse grid cells from being assigned weights. To handle this situation, structured spare weights are set. As a backup signal, it consists of a low vegetation cover index derived from POI point density or NDVI; simultaneously defining... The effective weights for entering the conservation normalization are:

[0097] (7)

[0098] in, For effective weights, As the initial weights, This is the scaling factor for the backup signal. It is a very small positive number. As a backup signal, For fine mesh element i, the coarse mesh element belongs to The sum of the initial weights. For fine mesh element i, the coarse mesh element belongs to The summation of internal backup signals.

[0099] When all surrogate variables within a coarse grid cell are zero, each effective 1km grid cell obtains the same result. After normalization, this is equivalent to uniform distribution. Missing values ​​and invalid observations are controlled by an effective mask, while true low-activity fine-grid cells are still preserved. In the middle, and when the basic agent is insufficient, it participates in the allocation through the backup weight.

[0100] Specifically, while rule-based allocators offer explicit interpretability, the spatial combination relationships between proxy variables are still manually defined. To enhance the expressive power of share estimation for complex spatial structures, this invention constructs a learning-based share allocator, denoted as MCSD-Net. This model does not directly predict absolute emissions, but rather predicts the share of each 1km fine-grid cell within its respective inventory grid cell; the predicted share is then multiplied by the total original inventory grid volume, thus the model cannot alter the total coarse-resolution emissions inventory volume.

[0101] The aligned spatial surrogate variables are organized into a multichannel feature tensor X:

[0102] (8)

[0103] Where C represents the number of input channels, and H and W represent the number of rows and columns of a 1km grid in the study area, respectively. Input channels include proxy variables such as roads, nighttime lights, land use area share, population, POI, and NDVI. Invalid fine grid cells are marked by a spatial mask and do not participate in normalization and loss calculation within coarse grid cells.

[0104] The MCSD-Net, a learning-based share allocator of this invention, takes a multi-channel surrogate variable tensor as input and outputs a single-channel response map corresponding to a 1km grid. This response map is not directly interpreted as emissions; instead, it is first converted into a non-negative response and then normalized to a learned share within the inventory grid cell.

[0105] The learning-based share allocator MCSD-Net uses a neural network model. Let the output of the neural network be... , where θ represents the neural network model parameters. The output is positive-valued to obtain a non-negative response, and then normalized to a learning share within the list grid cells.

[0106] (9)

[0107] The corresponding fine-grid cell emission rate is:

[0108] (10)

[0109] In the formula, The output response of the neural network model at fine grid cell i. For neural network model parameters, For multichannel surrogate variable tensors, The normalized share of fine grid cell i within its corresponding coarse grid cell, as output by the learning allocator. The set of effective fine mesh elements within the coarse mesh element to which fine mesh element i belongs. For the learning-type emissions of fine grid cell i, This represents the total number of coarse grid cells to which fine grid cell i belongs in the original inventory.

[0110] Equations 9 and 10 share the same conservation backend as the rule-based path, therefore the learning-based results also satisfy the total conservation of the inventory grid. This is relevant to the Chengdu-Chongqing Twin-City Economic Circle. Lacking a publicly available 1km reference field, MCSD-Net employs teacher-guided weakly supervised share learning. The conservation results generated by the rule-based allocator are first converted into shares within the list grid cells, which are then used as teacher shares.

[0111] (11)

[0112] In the formula, For list grid cells The proportion of regular teachers in inner fine grid cell i. For the initial weights Effective weights are formed using the aforementioned effective weight expression. Then, the regular emission amount of fine grid cell i is obtained by normalizing the total amount of coarse grid according to the conservation law. The emission rate is the regular emission rate of fine grid cell j within the same coarse grid cell.

[0113] The superscript (R) indicates a regular path. When the total teacher output within a coarse grid cell is 0, that coarse grid cell is not included in the corresponding training loss. It is a weakly supervised target obtained from spatial proxy variables and conservation normalization, not the actual observed 1km emission label.

[0114] The training loss of the above model uses the mean square error of the mask share on an effective 1km grid cell. :

[0115] (12)

[0116] Where, m i For effective spatial masking, This indicates whether the coarse grid cell to which the fine grid cell belongs has an effective teacher share. Training, validation, and testing are grouped and split by coarse grid cells to avoid fine grid cells within the same coarse grid cell entering different data subsets simultaneously, making the validation error closer to the model's generalization error on unknown coarse grid cells.

[0117] In practical implementation, the data involved in this invention are divided into two categories: emission inventory data provides coarse-resolution grid total constraints, and spatial proxy variable data provides emission spatial differences within a 1km grid. The product version, year, and spatiotemporal scope of the main datasets are shown in Table 1.

[0118] Table 1. Main datasets used in this invention and their applications

[0119]

[0120] The present invention Downscaling uses the EDGAR list as the primary input and incorporates the GRACED list for cross-list comparisons. EDGAR is maintained by the European Union Joint Research Centre, and this invention uses its 2019 publication. The total number of departments is used as the baseline input. GRACED provides global... The emission sites have a higher time resolution than conventional annual inventories; this invention uses its 12-month emission data from 2019. After the field data is aggregated into annual values, it is used to observe whether changes in inventory input alter the 1km share structure and extreme hotspot locations. To further observe the operational performance of MCSD on another type of pollutant, this invention also incorporates INTAC 2017 data. The list will include the emission targets from Turn to and continue to use The experiment included mapping between coarse and 1km grid cells, a proxy variable system, and a conservation normalization structure. Spatial proxy variables were used to describe differences in human activity intensity, traffic structure, and surface function within the inventory grid cells. Nighttime light, population, and POIs primarily reflected areal activity intensity. Nighttime light could identify urban built-up areas and high-activity areas, but may be affected by saturation effects in the urban core area; population data reflected the spatial clustering of residential activities and emissions; POI density supplemented information on urban functional nodes such as commercial facilities, public services, and transportation hubs. The road network characterized linear transportation corridors, providing spatial proxy for traffic-related emissions and emissions extensions along intercity corridors. Land use and NDVI primarily characterized the surface function background. The share of built-up land area reflected urban development intensity, while land types such as cultivated land, forest land, grassland, and water were used to distinguish built-up areas from surrounding agricultural and forestry spaces. NDVI reflected vegetation cover and could help identify low-construction-intensity areas. These variables, from the perspectives of activity intensity, transportation corridors, and surface background, constituted the multi-source spatial proxy variable inputs for MCSD. The study area location, auxiliary data, and grid alignment relationships are detailed in [link to relevant documentation]. Figure 3 .

[0121] MCSD maintains the total number of grid cells in both the EDGAR and GRACED inventory inputs. The downscaling results are then re-aggregated to... After gridding the list, the maximum absolute conservation deviations are respectively and All values ​​are within the floating-point calculation error range. The mapping between coarse grid cells and 1 km grid cells and the effective mask did not introduce any additional total deviation in actual calculations.

[0122] like Figure 4 As shown, 1 km under EDGAR input Emissions have formed two major high-value cores in the two main urban areas of Chengdu and Chongqing, extending towards the central region along intercity corridors within the economic circle and surrounding industrial clusters. Compared to the original... The list, when downscaled, reveals a clearer emission gradient between the city's edge, transportation corridors, and surrounding low-intensity development areas. High-value areas within the plains of City 1 expand in a sheet-like pattern, while the areas surrounding the main city of City 2 are constrained by topography and river valley development space, forming a strip-like clustering pattern.

[0123] like Figure 5As shown, the spatial morphology of the 1 km result under the GRACED input is highly similar to that of EDGAR, with high-emission areas still located in the dual core of City 1 and City 2, and the industrial and transportation corridors between the two cities. The difference lies in the fact that the total grid size of the GRACED inventory is significantly higher than that of EDGAR, resulting in an overall amplified absolute emission value per 1 km. The proxy variable determines the share within the inventory grid cell, while the coarse-resolution emission inventory determines the absolute emissions. The high degree of spatial similarity between the two inventories indicates that share allocation is mainly controlled by the proxy variable.

[0124] This invention is in In downscaling, MCSD maintained the total grid size conservation in both the EDGAR 2019 inventory and the GRACED 2019 comparative inventory. The downscaled spatial pattern restored the emission structure of the dual core of City 1 and City 2, as well as the industrial and transportation corridors between the two cities. Cross-inventory comparisons showed that the Pearson correlation coefficient for the effective 1 km grid share reached 0.99989, indicating a high degree of consistency in the fine-scale share structure under the two inventory inputs; however, extreme hotspots clustered in... Below the threshold, Jaccard drops to 0.431, but locally high values ​​are still influenced by the source of the inventory, departmental coverage, and activity data settings. (In INTAC) In migration applications, the main spatial patterns of rule-based and learning-based allocators are highly consistent. The adjustments made by learning-based allocators are concentrated within a few high-emission grid cells, and the total adjustment amount accounts for only a small percentage of the total emissions. This indicates that the learning-oriented path is mainly based on the regular spatial structure with limited modifications.

[0125] The aforementioned studies collectively demonstrate that within the Chengdu-Chongqing economic circle, the continuous built-up areas of urban plains, Chongqing's mountainous cities, intercity transportation corridors, and peripheral agricultural and forestry spaces are interspersed over short distances, making it difficult for coarse-grid unit mean values ​​to reflect the spatial differences in emissions within this urban cluster. MCSD, by expanding emission distribution to a 1 km scale while maintaining a constant total inventory grid size, effectively recovers fine-scale spatial structure using proxy variables in areas dominated by area sources. However, this recovery ability gradually weakens as the proportion of point sources increases, requiring the introduction of location and emission information of point source facilities in areas dominated by strong point sources. For areas lacking publicly available kilometer-level data... Referring to the western urban cluster, this invention provides a multi-faceted stability assessment of downscaling results through cross-inventory comparisons and cross-pollutant migration applications.

[0126] Example 2

[0127] like Figure 6As shown, the difference between this embodiment and Embodiment 1 is that this embodiment provides a spatiotemporal dynamic carbon emission accounting system driven by multimodal data fusion. This system corresponds one-to-one with the spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion in Embodiment 1. The system includes:

[0128] The acquisition unit is used to acquire raw emission inventory data and multiple proxy variables for the study area; the raw emission inventory data is multi-source low-resolution data.

[0129] The model building unit is used to build a carbon emission total model driven by multimodal data fusion. The carbon emission total model driven by multimodal data fusion includes a rule-based allocator and a learning allocator. The rule-based allocator is used to construct the initial weights inside the list grid cell based on the proxy variables. The learning allocator is used to learn the share structure inside the list grid cell from the proxy variables in the multi-channel space based on the initial weights.

[0130] The grid allocation unit is used to allocate the total carbon emission amount in the inventory grid based on the original emission inventory data and proxy variables, using a multimodal data fusion-driven total carbon emission model, to obtain high-resolution results.

[0131] The execution process of each unit can be carried out according to the steps of the spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion in Example 1, and will not be described in detail in this example.

[0132] Meanwhile, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion.

[0133] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0134] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0135] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0136] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0137] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion, characterized in that, The method includes: The original emission inventory data and multiple proxy variables for the study area are obtained; the original emission inventory data is multi-source low-resolution data. A multimodal data fusion-driven total carbon emission model is constructed. The multimodal data fusion-driven total carbon emission model includes a rule-based allocator and a learning allocator. The rule-based allocator is used to construct initial weights within the list grid cells based on the proxy variables. The learning allocator is used to learn the share structure within the list grid cells from the proxy variables in the multi-channel space based on the initial weights. Based on the original emission inventory data and proxy variables, the total carbon emission amount is allocated to the inventory grid using the multimodal data fusion-driven total carbon emission model, resulting in high-resolution results.

2. The spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion according to claim 1, characterized in that, The proxy variables include roads, nighttime lighting, construction land, population, POI, and NDVI.

3. The spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion according to claim 1, characterized in that, The rule-based allocator is an adaptive fusion coefficient based on the non-zero coverage of the proxy variable, which enables the weight allocation to be dynamically adjusted according to the proxy variable in different coarse grid cells to obtain the initial weight. The learning-based allocator employs a teacher-guided, weakly supervised learning strategy, using the initial weights as the teacher signal for training.

4. The spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion according to claim 3, characterized in that, Based on the adaptive fusion coefficients with non-zero coverage of proxy variables, the weight allocation can be dynamically adjusted according to the proxy variables within different coarse grid cells to obtain initial weights, including: Based on the surrogate variables, calculate the non-zero coverage of the first and second surrogate variables within the coarse grid cells; Based on the non-zero coverage of the first and second proxy variables, construct adaptive fusion coefficients; The fused signal of the two is calculated based on the adaptive fusion coefficient and the corresponding fine mesh cell proxy signal; The initial weights are calculated based on the fused signal and the third proxy variable.

5. The spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion according to claim 4, characterized in that, The expression for the initial weights is: ； ； ； ； In the formula, As the initial weights, For signal fusion, For the third surrogate variable share of fine grid cell i, The enhancement coefficient for the third proxy variable; The coarse mesh element to which fine mesh element i belongs The fusion coefficient, i.e., the adaptive fusion coefficient; For the first surrogate variable signal of fine mesh element i, This is the second proxy variable signal for the fine mesh element i; For adaptive fusion coefficients, This represents the non-zero coverage of the surrogate variable within the coarse grid cell c. `clip()` restricts the adaptive fusion coefficients to a given interval. for proxy variables Non-zero coverage within coarse grid cell c, for proxy variables Non-zero coverage within coarse grid cell c; This is the lower bound of the adaptive fusion coefficient. This represents the upper limit of the adaptive fusion coefficient; For non-zero coverage, This is the set of effective fine grid cells within coarse grid cell c after effective masking. The superscript v indicates the effective state and is used to exclude fine grid cells corresponding to outside the study area, missing values, or invalid observations. For indicator functions, Let z be the value of the surrogate variable on the fine mesh cell i.

6. The spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion according to claim 4, characterized in that, After obtaining the initial weights, the following is also included: The low vegetation cover index derived from the fourth and fifth proxy variables constitutes a backup signal; the initial weights are optimized and allocated based on the backup signal to obtain effective weights. The expression for the effective weight is: ； in, For effective weights, As the initial weights, This is the scaling factor for the backup signal. It is a very small positive number. As a backup signal, For fine mesh element i, the coarse mesh element belongs to The sum of the initial weights. The coarse mesh element to which fine mesh element i belongs The summation of internal backup signals.

7. The spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion according to claim 3, characterized in that, The teacher share in the teacher-guided, weakly supervised learning strategy is: ； In the formula, For the regular teacher share of fine grid cell i within list grid cell g(i), For the initial weights Effective weights are formed using the aforementioned effective weight expression. , The emission rate is the regular emission rate of fine grid cell j within the same coarse grid cell.

8. The spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion according to claim 1, characterized in that, The original emission inventory data includes the Global Atmospheric Research Emissions Database (EDGAR), the Global Gridded Daily Carbon Dioxide Emissions Dataset (GRACED), and the Integrated Emission Inventory of Air Pollutants in China (INTAC).

9. A spatiotemporal dynamic carbon emission accounting system driven by multimodal data fusion, characterized in that, The system includes: The acquisition unit is used to acquire raw emission inventory data and multiple proxy variables for the study area; the raw emission inventory data is multi-source low-resolution data. A model building unit is used to build a carbon emission total model driven by multimodal data fusion. The carbon emission total model driven by multimodal data fusion includes a rule-based allocator and a learning allocator. The rule-based allocator is used to construct initial weights within the list grid cells based on the proxy variables. The learning allocator is used to learn the share structure within the list grid cells from the proxy variables in the multi-channel space based on the initial weights. The grid allocation unit is used to allocate the total carbon emission amount to the grid based on the original emission inventory data and proxy variables, and to obtain high-resolution results.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the spatiotemporal dynamic carbon emission accounting method driven by multimodal data fusion as described in any one of claims 1 to 8.

Images